TLDR: A new AI framework addresses the challenge of predicting low-altitude network coverage, crucial for the growing drone economy. It overcomes issues like proprietary antenna data and sparse measurements by using expert knowledge to compress complex base station parameters and a specialized neural network that disentangles different aspects of signal propagation. This approach significantly improves prediction accuracy and generalizability, validated by real-world data showing a 7% error reduction and practical MAE errors at the 5 dB level, even with limited data or unknown base station types.
The burgeoning low-altitude economy, driven by innovations like unmanned aerial vehicles (UAVs), low-altitude logistics, and aerial surveillance, has created an urgent need for reliable and accurate Low-Altitude Network Coverage (LANC) prediction. Ensuring seamless connectivity in this unique airspace is vital for efficient network management and the design of aerial corridors.
Traditionally, network coverage prediction has focused on ground-level networks. However, low-altitude environments present distinct challenges. Unlike terrestrial settings, low-altitude networks often have fewer obstructions, leading to signals primarily from direct paths. This means that the actual coverage is heavily influenced by the multi-beam capabilities of Base Stations (BSs) and their collaborative effects, shifting the focus from environmental scattering to antenna configurations and beamforming patterns. A major hurdle is that antenna beam patterns, crucial for accurate forecasting, are typically proprietary and not easily accessible. Furthermore, collecting extensive low-altitude road test data is expensive, often resulting in only sparse data samples per BS, leading to imbalanced feature sampling and limited generalizability of prediction models.
To address these significant obstacles, a novel framework has been introduced. This framework employs a dual strategy: expert knowledge-based feature compression and disentangled representation learning.
Expert Knowledge for Smarter Data
The first part of the strategy involves leveraging communications expertise to compress the complex operational parameters of Base Stations. These parameters include details like the BS’s absolute location, static characteristics of its antenna beam (e.g., AAU type, number of channels, carrier frequency), orientation characteristics (like horizontal azimuth and mechanical down tilt), and additive signal strength features (transmission power, bandwidth). By applying expert knowledge, these high-dimensional parameters are transformed into more balanced and compressed features, such as relative angles and distances between the BS and the target location. This process effectively tackles the issue of imbalanced feature sampling, providing a robust foundation for accurate and efficient LANC prediction.
Disentangled Learning for Better Predictions
The second key component is a disentangled representation learning network. This network is designed to learn the latent representations of distinct, independent factors influencing signal propagation. Inspired by how signals propagate, the framework separates the prediction task into components related to distance fading, frequency fading, and antenna gain. Each of these components is handled by a dedicated sub-network, typically a Multi-Layer Perceptron (MLP). By disentangling these factors, the model can better understand and aggregate the semantic representations of latent features, significantly enhancing its generalizability and interpretability, especially when data samples are limited. The outputs from these separate networks are then combined to produce the final coverage prediction.
Also Read:
- AI Model Enhances UAV Flight State Recognition
- Intelligent Route Planning for Urban Air Mobility Ride-Sharing
Real-World Validation and Practical Impact
The efficacy of this framework was rigorously evaluated through experiments using real-world low-altitude data collected in Nanchang, Jiangxi Province, China. The results were highly promising: the framework achieved a 7% reduction in error compared to the best baseline algorithm. Ablation studies further demonstrated that the proposed framework reduced the mean absolute error (MAE) by approximately 20% compared to traditional deep neural networks. In practical scenarios, the framework achieved prediction accuracy with MAE errors at the 5 dB level, confirming its reliability and applicability.
The framework also proved robust in scenarios with low data availability, such as early stages of low-altitude network planning or when applying the algorithm across different cities or operators with few matched antenna types. It can predict coverage for individual base stations, identify multi-BS coverage holes, and even adapt when certain BS parameters, like the AAU type, are unknown. For instance, it successfully predicted coverage in Ganzhou, Jiangxi Province, where BSs were from a different manufacturer than the training data, demonstrating its transferability.
This research marks a significant advancement in LANC prediction, offering a data-driven model that innovatively leverages BS operational parameters. By combining expert knowledge-based feature compression with a model-guided disentangled representation network, it provides a powerful tool for designing aerial corridors and managing the rapidly expanding low-altitude economy. For more details, you can refer to the full research paper here.
